Face detection and recognition is one the area where Deep Learning is incredibly useful. There are many studies and datasets related to human faces and their detection/recognition. In this article we will implement Machine Learning pipeline for face detection and recognition using few libraries and CNN model.

Last time we used ANN to train a Deep Learning model for image recognition using MNIST dataset. This time we are going to look at more advanced network called Convolutional Neural Network or CNN in short.

CNN is designed to tackle image recognition problem. However, it can be used not only for image recognition. As we have seen last time, ANN using just hidden layers can learn quite well on MNIST. However, for real life use cases we need higher accuracy. The main idea of CNN is to learn how to recognise object in their different shapes and positions using specific features of the image data. The goal of CNN is better model regularisation by using convolution and pooling operations.

Deep Feed Forward Neural Network is one of the type of Artificial Neural Networks, which is also able to classify computer images. In order to feed pixel data into the neural net in RBG/Greyscale/other format one can map every pixel to network inputs. That means every pixel becomes a feature. It may sound scary and highly inefficient to feed, let's say, 28 hieght on 28 width image size, which is 784 features to learn from. However, neural networks can learn from the pixel data successfully and classify unseen data. We are going to prove this.

Please note, there are additional type of networks which are more efficient in image classification such as Convolutional Neural Network, but we are going to talk about that next time.

Dataset

Adam is one more optimization algorithm used in neural networks. It is based on adaptive estimates of lower-order moments. It has more hyper-parameters than classic Gradient Descent to tune externally

Good default settings for the tested machine learning problems are:

• α = 0.001, // learning rate. We have already seen this one in classic Gradient Descent.
• β₁ = 0.9,
• β₂ = 0.999
• eps = 10−8.

In this article we are going to use Scala mini-library for Deep Learning that we developed earlier in order to study basic linear regression task. We will learn model weights using perceptron model, which will be our single unit network layer that emits target value. This model will predict a target value yHat based on two trained parameters: weight and bias. Both are scalar numbers. Weights optimization is going to be based on implemented Gradient descent algorithm:

• Gradient Descent Optimization
• Training Loop

Model equation:

y = bias + weight * x